US20050125307A1

US20050125307A1 - Approach for estimating user ratings of items

Info

Publication number: US20050125307A1
Application number: US10/773,843
Authority: US
Inventors: Neil Hunt; Stanley Lanning; W. Hastings; Shawn Purcell; Lawrence Shih; John Ciancutti
Original assignee: Netflix Inc
Current assignee: Netflix Inc
Priority date: 2000-04-28
Filing date: 2004-02-06
Publication date: 2005-06-09
Also published as: EP1730686A2; US20100010877A1; CA2556026A1; JP2007528060A; WO2005076890A3; US8271338B2; US7617127B2; AU2005213374A1; WO2005076890A2

Abstract

According to another embodiment of the invention, an approach is provided for estimating how a particular user would rate a particular item from a plurality of items. The approach is applicable to any type of items, including rental items such as movies, music and games, and the invention is not limited to any particular type of item. One or more items from the plurality of items that have ratings similar to the particular item are identified. Then, one or more other users are identified that have given ratings to the one or more items that are substantially similar to ratings given by the particular user to the one or more items. Finally, an estimation is made of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users.

Description

RELATED APPLICATIONS

This application is a continuation-in-part of and claims priority to U.S. patent application Ser. No. 09/884,816 (Attorney Docket No. 56055-0013) filed on Jun. 18, 2001, the entire contents of which are incorporated herein by reference in their entirety for all purposes, which is a continuation-in-part of U.S. patent application Ser. No. 09/561,041, now issued as U.S. Pat. No. 6,584,450 (Attorney Docket No. 56055-0011), filed on Apr. 28, 2000 and also claims priority to U.S. Provisional Patent Application No. 60/212,193 (Attorney Docket No. 56055-0012) filed on Jun. 16, 2000 and U.S. Provisional Patent Application No. 60/244,793 (Attorney Docket No. 56055-0014) filed on Oct. 31, 2000.

FIELD OF THE INVENTION

The present invention relates to inventory rental, and more specifically, to an approach for estimating user ratings of items.

BACKGROUND OF THE INVENTION

Conventional inventory rental models are typically based upon renting items for fixed rental periods and charging late fees for keeping rented items beyond a specified return date. As used herein, the term “item” refers to any item of commerce. These types of inventory models suffer from several significant limitations. First, conventional rental models require customers to make the decision of what items to rent at substantially the same time as the decision of when to rent the items. An example that illustrates this limitation is a video rental business. Customers go to a video rental store and select particular movies to rent at that time. The customers take the movies home and must return them by a particular due date or be charged a late fee. In this situation, the customers cannot decide what movies to rent before actually renting them. The customers may have a particular movie in mind, but there is no guarantee that the video rental store has the particular movie in stock. Moreover, due dates are inconvenient for customers, particularly for “new release” movies that are generally due back the next day.
Given the current demand for inventory rental and the limitations in the prior approaches, an approach for renting items to customers that does not suffer from limitations associated with conventional inventory rental models is highly desirable. In particular, an approach for renting inventory items to customers that allows separation of customers' decisions of what items to rent from when to rent the items is highly desirable.
There is a further need for an approach for renting items to customers on a continuous basis that avoids the use of fixed due dates or rental “windows” appurtenant to conventional rental models.
There is yet a further need for an approach for renting movies, games and music to customers that is more convenient and flexible to customers than conventional approaches.
There is also a need for an approach for estimating how a user would rate an item that the user has not yet rated.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the invention are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which like reference numerals refer to similar elements and in which:
FIG. 1 is a diagram depicting an approach for renting items to customers according to an embodiment.
FIG. 2 is a flow diagram depicting an approach for renting items to customers according to an embodiment.
FIG. 3 is a flow diagram depicting a “Max Out” approach for renting items to customers according to an embodiment.
FIG. 4 is a flow diagram depicting a “Max Turns” approach for renting items to customers according to an embodiment.
FIG. 5 is a diagram depicting an approach for renting audio/video items to customers over the Internet according to an embodiment.
FIG. 6 is a flow diagram illustrating an approach for renting audio/video items to customers over the Internet using both “Max Out” and “Max Turns” according to an embodiment.
FIG. 7 is a flow diagram that depicts an approach for estimating how a user would rate an item according to an embodiment of the invention.
FIG. 8 is a block diagram that depicts users' ratings of items arranged in an array.
FIG. 9 is a block diagram that depicts an arrangement for recommending items according to an embodiment of the invention.
FIG. 10 is a block diagram that depicts another arrangement for recommending items according to an embodiment of the invention.
FIG. 11 is a block diagram of a computer system upon which embodiments of the invention may be implemented.

DETAILED DESCRIPTION OF THE INVENTION

In the following description, for the purposes of explanation, specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent that the invention may be practiced without these specific details. In other instances, well-known structures and devices are depicted in block diagram form in order to avoid unnecessarily obscuring the invention.
Various aspects and features of example embodiments of the invention are described in more detail hereinafter in the following sections:

- I. FUNCTIONAL OVERVIEW
- II. ITEM SELECTION CRITERIA
- III. ITEM DELIVERY
- IV. “MAX OUT”
- V. “MAX TURNS”
- VI. INVENTORY MANAGEMENT
- VII. ITEM RECOMMENDATION
- VIII. IMPLEMENTATION MECHANISMS
  I. Functional Overview

FIG. 1 is a block diagram 100 that illustrates an approach for renting items to customers according to various embodiments described herein. As used herein, the term “items” refers to any commercial goods that can be rented to customers. Examples of items include movies, music and games stored on a non-volatile memory such as a tape, other magnetic medium, optical medium, read-only memory or the like, and the invention is not limited to any particular type of item. In general, the decision of what items to rent is separated from the decision of when to rent the items. Customers may specify what items to rent using one or more item selection criteria separate from deciding when to receive the specified items. Furthermore, customers are not constrained by conventional rental “windows” and instead can have continuous, serialized rental of items.
According to one embodiment, a customer 102 provides one or more item selection criteria to a provider 104 over a link 106. Link 106 may be any medium for transferring data between customer 102 and provider 104 and the invention is not limited to any particular medium. Examples of link 106 include, without limitation, a network such as a LAN, WAN or the Internet, a telecommunications link, a wire or optical link or a wireless connection.
The item selection criteria indicate items that customer 102 desires to rent from provider 104. In response to receiving the item selection criteria from customer 102, provider 104 provides the items indicated by the item selection criteria to customer 102 over a delivery channel 108. Delivery channel 108 may be implemented by any mechanism or medium that provides for the transfer of items from provider 104 to customer 102 and the invention is not limited to any particular type of delivery channel. Examples of delivery channel 108 include, without limitation, mail delivery, courier delivery or delivery using a delivery agent. Provider 104 may be centralized or distributed depending upon the requirements of a particular application.
According to an embodiment, a “Max Out” approach allows up to a specified number of items to be rented simultaneously to customer 102 by provider 104. According to another embodiment, a “Max Turns” approach allows up to a specified number of item exchanges to occur during a specified period of time. The “Max Out” and “Max Turns” approaches may be used together or separately with a variety of subscription methodologies.
The approach just described for renting items to customers is now described with reference to a flow diagram 200 of FIG. 2. After starting in step 202, in step 204, customer 102 creates item selection criteria. In step 206, customer 102 provides the item selection criteria to provider 104. In step 208, in response to provider 104 receiving the item selection criteria from customer 102, provider 104 provides one or more items indicated by the item selection criteria to customer 102. The process is complete in step 210.
II. Item Selection Criterial
The one or more item selection criteria provided by customer 102 to provider 104 indicate the particular items that customer 102 desires to rent from provider 104. Thus, the item selection criteria define a customer-specific order queue that is fulfilled by provider 104. According to one embodiment, the item selection criteria specify attributes of items to be provided by provider 104 to customer 102. Item selection criteria may specify any type of item attributes and the invention is not limited to particular item attributes. Examples of item attributes include, without limitation, identifier attributes, type attributes and cost attributes. Item selection criteria may be changed at any time to reflect changes in items that customers desire to rent from a provider.
III. Item Delivery
According to one embodiment, items are delivered by provider 104 to customer 102 over delivery channel 108 based upon item delivery criteria. More specifically, the delivery of items from provider 104 to customer 102 is triggered by item delivery criteria being satisfied. The item delivery criteria may include a wide range of criteria and the invention is not limited to any particular item delivery criteria. Examples of item delivery criteria include, without limitation, customer request/notification, customer notification that an item is being returned, customer return of an item, the occurrence of a specified date, the elapsing of a specified period of time or a customer payment.
The item delivery criteria may be specified by customer 102 to provider 104 or negotiated by customer 102 and provider 104 as part of a subscription service. For example, a particular subscription service may include item delivery criteria that specifies that a particular number of items are to be delivered monthly. As another example, item delivery criteria may specify that an initial set of items is to be delivered by provider 104 to customer 102 upon initiation of a subscription service and that additional items are to be delivered to customer 102 upon return of items to provider 104. Item delivery criteria may be applied uniformly to all items to be delivered to a customer, or may be item specific. For example, item delivery criteria may specify a particular date, i.e., the third Wednesday of every month, for all item deliveries. Alternatively, separate item delivery dates may be assigned to each item.
IV. “Max Out”
According to one embodiment, a “Max Out” approach is used to manage the number of items that may be simultaneously rented to customers. According to the “Max Out” approach, up to a specified number of items may be rented simultaneously to a customer. Thus, the “Max Out” approach establishes the size of an inventory of items that may be maintained by customers. The specified number of items may be specific to each customer or may be common to one or more customers. In the present example, if the specified number of items is three, then up to three items may be rented simultaneously by provider 104 to customer 102. If the specified number of items are currently rented to customer 102 and the specified item delivery criteria triggers the delivery of one or more additional items, then those items are not delivered until one or more items are returned by customer 102 to provider 104.
According to one embodiment, in situations where the specified number of items are currently rented to customer 102 and the specified item delivery criteria triggers the delivery of one or more additional items, then the one or more additional items are delivered to customer 102 and customer 102 and a surcharge is applied customer 102. The specified number of items may then be increased thereafter to reflect the additional items delivered to customer 102 and increase the size of the inventory maintained by customer 102. Alternatively, the specified number of items may remain the same and number of items maintained by customer 102 returned to the prior level after items are returned to provider 104 by customer 102. When used in conjunction with the “Max Turns” approach described hereinafter, the specified number of items may be unlimited.
The “Max Out” approach for managing the number of items that may be simultaneously rented to customers is now described with reference to a flow diagram 300 of FIG. 3. After starting in step 302, in step 304, one or more initial items are delivered to customer 102 to establish the inventory maintained by customer 102. Note that an initial delivery of items is not required and according to one embodiment, the inventory of customer 102 is incrementally established over time.
In step 306, a determination is made whether the item delivery criteria have been satisfied. If not, then the determination continues to be made until the item delivery criteria are satisfied. As described previously herein, the delivery criteria may include customer notification generally, customer notification that an item is being returned, the actual return of an item, the occurrence of a specific date, or that a specified amount of time has elapsed.
Once the item delivery criteria are satisfied, then in step 308, a determination is made whether the specified number of items have been delivered. If not, then control returns to step 304 and one or more additional items are delivered by provider 104 to customer 102. If however, in step 308, the specified number of items have been delivered, then in step 310, a determination is made whether the specified number of items, i.e., the “Max Out” limit, is to be overridden. As previously described, the specified number of items may be overridden by increasing the specified number of items, i.e., the “Max Out” limit, to allow additional items to be delivered to customer 102 and charging a fee to customer 102. Alternatively, the specified number of items is not changed and a surcharge applied to customer 102. This process continues for the duration of the subscription and is then complete in step 312.
V. “Max Turns”
According to one embodiment, a “Max Turns” approach is used to rent items to customers. According to the “Max Turns” approach, up to a specified number of item exchanges may be performed during a specified period of time. For example, referring to FIG. 1, suppose that provider 104 agrees to rent items to customer 102 with a “Max Turns” limit of three items per month. This means that customer 102 may make up to three item exchanges per month. This approach may be implemented independent of the number of items that a customer may have rented at any given time under the “Max Out” approach. The approach is also independent of the particular item delivery criteria used.
According to one embodiment, the “Max Turns” approach is implemented in combination with the “Max Out” approach to rent items to customers. In this situation, up to a specified number of total items are simultaneously rented to customer 102 and up to a specified number of item exchanges may be made during a specified period of time. Thus, using the “Max Out” and the “Max Turns” approaches together essentially establishes a personal item inventory for customer 102 based upon the “Max Out” limit that may be periodically refreshed based upon the “Max Turns” limit selected.
In some situations, customer 102 may wish to exchange more than the specified number of items during a specified period. According to one embodiment, in this situation, provider 104 agrees to rent additional items above the specified number to customer 102 and to charge customer 102 for the additional items. For example, suppose that provider 104 agrees to rent items to customer 102 with up to three item turns (exchanges) per month. If, in a particular month, customer 102 requires two additional turns, then the two additional items are provided to customer 102 and a surcharge is applied to customer 102 for the additional two items.
In other situations, customer 102 may not use all of its allotted turns during a specified period. According to one embodiment, customers lose unused turns during a subscription period. For example, if customer 102 has a “Max Turns” limit of four item exchanges per month and only makes two item exchanges in a particular month, then the two unused exchanges are lost and cannot be used. At the start of the next month, customer 102 would be entitled to four new item exchanges.
According to another embodiment, customers are allowed to carry over unused turns to subsequent subscription periods. For example, if customer 102 has a “Max Turns” limit of four item exchanges per month and only makes two item exchanges in a particular month, then the two unused exchanges are lost and cannot be used. At the start of the next month, customer 102 would be entitled to six new item exchanges, two from the prior month and four for the current month.
The “Max Turns” approach for renting items to customers is now described with reference to a flow diagram 400 of FIG. 4. After starting in step 402, in step 404, customer 102 and provider 104 agree upon the terms of the “Max Turns” agreement. Specifically, customer 102 and provider 104 agree at least upon the maximum number of turns that are allowed in a specified period of time.
In step 406, in response to one or more item delivery criteria being satisfied, provider 104 provides one or more items to customer 102 over delivery channel 108. Any type of item delivery criteria may be used with the “Max Turns” approach and the invention is not limited to any particular delivery criteria. For example, the initial one or more items may be delivered to customer 102 in response to a subscription payment made by customer 102 to provider 104, the initiation of a specified subscription period, or by request of customer 102 for the initial-rental items. The number of initial one or more items must not exceed the terms of the “Max Turns”agreement.
In step 408, in response to one or more delivery criteria being satisfied, a determination is made whether additional items can be provided to customer 102 within the terms of the “Max Turns” agreement. For example, if the number of items rented to customer in the current subscription period is less than the agreed-upon “Max Turns,” then additional items can be rented to customer 102 within the terms of the “Max Turns” agreement. In this situation, this determination may be made in response to customer 102 returning one or more items to provider 104, or by customer 102 requesting additional items.
If, in step 408, a determination is made that additional items can be rented to customer 102 within the terms of the “Max Turns” agreement, then control returns to step 406 where one or more additional items are rented to customer 102. If however, in step 408, a determination is made that additional items cannot be rented to customer 102 within the-terms of the “Max Turns” agreement, then in step 410, a determination is made whether to override the current agreement terms. If so, then in step 412, the agreement terms are changed to allow for a larger number of terms and customer 102 is charged accordingly, or the terms are left unchanged and a surcharge is applied for the additional items to be delivered. Control then returns to step 406, where one or more additional items are delivered to customer 102.
If in step 410, a determination is made that the current agreement is not to be overridden, then in step 414, no items are delivered to customer 102 until the next subscription period. For example, the request for additional items may be received at the end of a subscription period and instead of renting the additional items immediately, they are instead delivered during the subsequent subscription period. Control then returns to step 406 where one or more additional items are rented to customer or the process is complete in step 416.
The approach for renting items described herein is now described in the context of renting to customers audio/video (A/V) items, such as movies, games and music, stored on various media. FIG. 5 is a diagram 500 that depicts a set of customers 502 that desire to rent A/V items from a provider 504. Customers 502 communicate with provider 504 over links 506, the global packet-switched network referred to as the “Internet,” and a link 510.
Links 506 and 510 may be any medium for transferring data between customers 502 and the Internet 508 and between the Internet 508 and provider 504, respectively, and the invention is not limited to any particular medium. In the present example, links 506 and 510 may be connections provided by one or more Internet Service Providers (ISPs) and customers 502 are configured with generic Internet web browsers. Links 506 and 510 may be secure or unsecured depending upon the requirements of a particular application.
In accordance with an embodiment, customers 502 enter into a rental agreement with provider 504 to rent audio/video (A/V) items 512 from provider 504 according to the “Max Out” and/or “Max Turns” approaches described herein. The invention is not limited to any particular approach for entering into the rental agreement. For example, customers 502 and provider 504 may enter into a rental agreement by mail, telephone or over the Internet, by customers 502 logging into a web site associated with provider 504.
Customers 502 create and provide item selection criteria to provider 504 over links 506 and 510 and the Internet 508. The invention is not limited to any particular approach for specifying and providing item selection criteria to provider 504. For example, according to one embodiment, customers 502 provide item selection criteria to provider 504 in one or more data files. According to another embodiment, customers 502 log onto a web site of provider 504 and use a graphical user interfaced (GUI) to specify attributes of the movies and music that customers desire to rent from provider 504.
The item selection attributes may include any attributes that describe, at least in part, movies, games or music that customers 502 desire to rent. For movies, example attributes include, without limitation, title, category, director name, actor name and year of release. For games, example attributes include, without limitation, title and category. For music, example attributes include, without limitation, title, category, artist/group name and year of release. Customers 502 may identify specific movies or music by the item selection criteria, or may provide various attributes and allow provider 504 to automatically select particular movies and music that satisfy the attributes specified. For example, customers 502 may specify item selection criteria that include horror movies released in 1999 and let provider 504 automatically select horror movies that were release in 1999. As another example, customers 502 may specify item selection criteria that include adventure movies starring Harrison Ford. Customers 502 may also specify an order or priority for the specified item selection criteria. For example, customers 502 may specify specific movie titles and the order in which they want to receive them. As another example, customers 502 may specify that they want to receive a particular number of movies of different types.
Once customers 502 and provider 504 have entered into a rental agreement and customers 502 have provided item selection criteria to provider 504, then A/V items 512 are rented to customers 502 over delivery channels 514 in accordance with the terms of the rental agreement. Specifically, according to the “Max Out” approach described herein, an initial set of A/V items 512, such as movies, games and music, are delivered to customers 502 over delivery channels 514 according to the terms of the rental agreement. Subsequent A/V items 512 are delivered whenever the specified item delivery criteria are satisfied. For example, additional A/V items 512 may be delivered upon the return of one or more A/V items 512 to provider, a request from customers 502, the arrival of a particular date, e.g., a specific day of the month, or the expiration of a specified period of time, e.g., fifteen days.
In accordance with the “Max Out” approach described herein, once the maximum number of A/V items 512 have been rented to a particular customer 502, then no additional A/V items 512 are rented until one or more rented A/V items 512 are returned to provider 504, or unless a surcharge is applied to the particular customer 502. Alternatively, the rental agreement between the particular customer 502 and provider 504 may be modified to increase the maximum number of A/V items 512 that may be rented simultaneously to the particular customer 502.
The rental agreement between customers 502 and provider 504 may also specify a maximum number of turns in combination with the “Max Out” approach. In this situation, a maximum number of turns restricts how quickly customers 502 may refresh their A/V item 512 inventories. For example, suppose that a particular customer 502 agrees with provider 504 to rent up to four movies with a maximum of four turns per month. Under this agreement, the particular customer 502 may maintain a personal inventory of up to four movies and rent four new movies per month. Thus, the particular customer 502 can completely “replace” his personal inventory once per month. If the particular customer 502 agreed to a maximum of up to eight turns per month, then the particular customer 502 would be able to completely replace his personal inventory twice per month.
Provider 504 is illustrated as a single entity for purposes of explanation only. Provider 504 may be centralized or distributed depending upon the requirements of a particular application. For example, provider 504 may be a central warehouse from which all A/V items 512 are provided. Alternatively, provider 504 may be implemented by a network of distributed warehouses.
FIG. 6 is a flow diagram that illustrates an approach for renting A/V items 512, e.g., movies, to customers over a communications network such as the Internet using both “Max Out” and “Max Turns” according to an embodiment. Referring also to FIG. 5, after starting in step 602, in step 604, a customer 502 enters into a rental agreement with provider 504. In the present example, customer 502 uses a generic web browser to access an Internet web site associated with provider 504 and enter into a rental agreement that specifies that customer 502 may maintain a personal inventory of four movies (“Max Out” of four) and receive up to four new movies per month (“Max Turns” of four). Furthermore, the rental agreement specifies that new movies will be delivered upon return of a rented movie from customer 502, i.e., the delivery criteria is a return of a movie by the customer.
In step 606, customer 502 creates and provides movie selection criteria to provider 504 that indicates movies that customer 502 desires to rent. For example, the movie selection criteria may specify particular movie titles that customer 502 desires to rent. The movie selection criteria may also specify an order or priority in which customer 502 wishes to rent the movies. Instead of identifying particular movie titles, the movie selection criteria may specify movie preferences for customer 502, e.g., types of movies, directors, actors, or any other movie preferences or attributes. In this situation, provider 504 automatically selects particular titles that satisfy the movie selection criteria. For example, the movie selection criteria may specify a preference for action movies starring a particular actor, with a preference for “new release” movies. Provider 504 attempt to provide movies to customer 502 that best satisfy the preferences indicated by the movie selection criteria.
In step 608, one or more initial movies 512 are delivered to customer 502 over delivery channel 514. The one or more initial movies 512 may be delivered to customer 502 via mail, courier, delivery agent or any other suitable means negotiated between customer 502 and provider and the invention is not limited to any particular type of delivery mechanism. For purposes of explanation only, is presumed in the present example that movies are mailed between customer 502 and provider 504.
The one or more initial movies 512 establish the personal movie inventory of customer 502. Customer 502 may choose to receive any number of movies up to the “Max Out” limit of four movies. Typically, customer 502 will choose to initially receive four movies in the initial delivery.
Once the one or more initial movies 512 have been mailed to customer 502, then in step 610, a determination is made whether any movies 512 have been returned by customer 502 to trigger another movie delivery. In the present example, the delivery of additional movies is triggered by the receipt, e.g., via mail, of one or more movies from customer 502. In the situation where customer 502 elects to not receive the maximum number of movies 512 in the initial delivery, then the delivery of additional movies 512 may also be triggered by a request from customer 502 for additional movies 512. For example, customer 502 may notify provider 504 via telephone, email or by accessing the web site associated with provider 504.
If, in step 610, a determination is made that one or more movies 512 were received from customer 502, then in step 612, a determination is made whether the maximum number of turns (“Max Turns”) limit has been reached for the current cycle. In the present example, a determination is made whether four or more movies have been mailed in the current month. If not, then control returns to step 608, where one or more additional movies 512 are mailed to customer 502 via delivery channel 514 up to the “Max Out” limit of four.
If, in step 612, a determination is made that the “Max Turns” limit has been met for the current cycle, i.e., in the present example, four movies 512 have been mailed to customer 502 in the current month, then in step 614 a determination is made whether to override the current “Max Turns” limit. If so, then in step 616, a surcharge is applied to customer 502 and control returns to step 608 where the additional movies 514 are mailed to customer 502. If not, then in step 618, a determination is made whether to continue the subscription service. If so, then no additional movies are mailed to customer 502 during the current cycle, e.g., the current month, and the control returns to step 610. If, in step 618, a determination is made that service is not to be continued, then the process is complete in step 620.
In some situations, customer 502 may desire to increase or decrease the size of customer's 502 personal movie inventory by changing the current “Max Out” limit. According to one embodiment, customer 502 notifies provider 504, e.g., by telephone, mail, email or by accessing the web site associated with provider 504, that customer 502 wishes to change the “Max Out” limit. The movie rental agreement between customer 502 and provider 504 is then modified to reflect the change of the “Max Out” limit. In the situation where the “Max Out” limit is increased, then additional movies 512 may be immediately mailed to customer 502.
VI. Inventory Management
The approach described herein for renting items to customers provides superior inventory management to prior approaches. Specifically, the use of item selection criteria provides for efficient inventory management by allowing the greatest number of items to be rented at any given time. Moreover, the greatest number of customers are provided with their most preferred items. For example, customers may specify priorities for the items indicated by the item selection criteria. Thus, if a particular customer's first choice is not available, or already rented, then the item having the next highest priority can be rented to the particular customer. According to one embodiment, customers may indicate items that are not yet available for rent. Then, the items are delivered to customers when they become available.
For example, referring again to FIG. 5, suppose that a particular customer 502 desires to rent an as-yet-unreleased movie entitled “ABC.” The particular customer 502 indicates this movie to provider 504 by the item selection criteria. Since the movie ABC is not yet available, it cannot be delivered to the particular customer 502. However, when the movie ABC does become available, it can be shipped immediately to the particular customer 502, as well as other customers 502 who may have also requested the movie. This allows provider 504 to maximize the number of items rented while ensuring that customers 502 are able to rent the highest priority items that they requested.
According to another embodiment, as yet unknown items may also be rented by specifying attributes of the unknown items. For example, the particular customer 502 may request to rent the next new movie of a particular director, for which the exact name is unknown to the particular customer. As another example, the particular customer 502 may request to rent the next album of a particular group that is currently in process and does not yet have a title.
VII. Item Recommendation
A. Overview
In some situations, it is desirable to know how a user would rate an item. For example, in the context of recommending items to users, it is useful to know how a user would rate an item, so that a decision can be made to either recommend or not recommend the item to the user. A recommendation is more valuable to the user if it is likely that the user would rate the item highly because the user is more likely to be happy with the recommendation. An approach is provided for estimating how a user would rate an item. The approach generally involves estimating how a user would rate an item based upon similarities between items and similarities of ratings given to items by both the user and one or more other users. The approach is applicable to any type of item, for example rental items such as movies and games, and the invention is not limited to any particular type of item. The approach is applicable to a wide variety of contexts and may be used, for example in a recommendation engine to recommend items to users.
FIG. 7 is a flow diagram 700 that depicts an approach for estimating how a user would rate a particular item according to an embodiment of the invention. In step 702, one or more items from a plurality of items are identified that have ratings similar to the particular item. In step 704, one or more other users are identified that have given ratings to the one or more other items that are substantially similar to ratings given by the user to the one or more other items, and have also rated the particular item. In step 706, an estimation of how the user would rate the particular item is generated based upon ratings given by the one or more other users for the particular item. Each of these steps is described in more detail hereinafter.
FIG. 8 is a block diagram that depicts users' ratings of items arranged in an array 800. Array 800 includes columns of items 802 and rows of users 804. Each cell may be empty or contain data indicating that the user has not provided a rating for the item. For example, a specified value, such as a null value or other value, may be used to indicate that a user has not provided a rating for an item. Alternatively, each cell may contain data that represents a user rating of an item. Different types of data and values may be used to indicate a user rating of an item, depending upon the requirements of a particular application, and the invention is not limited to any particular data types or values. For example, ratings may be represented by numerical data, e.g., in a range of 1 to 5 that corresponds to a 5-star rating scale or binary data that corresponds to a favorable or unfavorable rating. Embodiments of the invention are described herein in the context of users' ratings in array 800 for explanation purposes only and the invention is not limited to users' ratings being maintained in any particular format or data structure. Furthermore, the relative size of the items 802 and users 804 dimensions depicted in FIG. 8 is not meant to imply that there must be more users 804 than items 802. Embodiments of the invention are described herein in the context of generating an estimated rating for a particular item 806, from items 802, and a particular user 808, from users 804. The estimated rating for the particular item 806 and particular user 808 is indicated in FIG. 8 by reference numeral 810.
B. Identifying Other Items that have Ratings Similar to the Particular Item
A set of similar items 812 that have ratings similar to the particular item 806 is identified. Similar items 812 are depicted in FIG. 8 as being located adjacent to the particular item 806 in array 800 for explanation purposes only. Similar items 812 may include any of items 802. A variety of techniques may be used to determine similar items 812, depending upon the requirements of a particular application, and the invention is not limited to any particular technique.
According to one embodiment of the invention, similar items 812 are items from items 802 that satisfy item similarity criteria with respect to particular item 806. Thus, other items that satisfy the item similarity criteria with respect to the particular item are considered to be similar to the particular item. Similarity is not necessarily symmetrical. Thus, although an item A may be considered similar to particular item 806 and therefore included in similar items 812 for particular item 806, particular item 806 may, or may not be, sufficiently similar to item A to be considered similar to item A and included in a set of similar items for item A. The number of items in similar items 812 may be limited to a specified number to reduce the amount of computational resources and time required to generate an estimated rating for particular item 806. For example, the specified number of items in similar items 812 may represent the first specified number of items that satisfied the item similarity criteria for particular item 806, or may represent the specified number of items that satisfied the item similarity criteria for particular item 806 with similarity above some specified level or threshold.
The item similarity criteria may include a wide variety of criteria, depending upon the requirements of a particular implementation, and the invention is not limited to any particular item similarity criteria. One example of item similarity criteria is a statistical correlation threshold. Thus, two items that satisfy the statistical correlation threshold are considered to be similar to each other. A wide variety of statistical measures may be used for this purpose, depending upon the requirements of a particular implementation, and the invention is not limited to any particular statistical measure of similarity. For example, a Pearson Correlation Coefficient may be used as the statistical measure of similarity as shown in equation (1) below: $\begin{matrix} r = \frac{n \sum X_{i} Y_{i} - \sum X_{i} \sum Y_{i}}{\sqrt{[n \sum X_{i}^{2} - {(\sum X_{i})}^{2}] [n \sum Y_{i}^{2} - {(\sum Y_{i})}^{2}]}} & (1) \end{matrix}$
where X_iare the ratings for one item or user and Y_iare the ratings for the other item or user, and the summation is applied across all co-ratings. As another example, a vector cosine distance may be used as the statistical measure of similarity as shown in equation (2) below: $\begin{matrix} r = \frac{\sum X_{i} Y_{i}}{\sqrt{\sum X_{i}^{2}} \sqrt{\sum Y_{i}^{2}}} & (2) \end{matrix}$
where X_iand Y_iare the ratings for one item or user.
Table I below contains example user ratings for items X and Y, the product of the user ratings for X and Y, the square of the user ratings for X and Y and the summations of these values:+

TABLE I

X Y X · Y X**2 Y**2

1 2 2 1 4

4

4 1 4 16 1

5 4 20 25 16

3 3 9 9 9

3

5 4 20 25 16

1

4 4 16 16 16

4 5 20 16 25

Summations 26 23 91 108 87
Note that the summations of the X and Y columns include only ratings for users who rated both item X and Y. $The Pearson Correlation Coefficient is$ $\frac{7 \times 91 - 26 \times 23}{\sqrt{[7 \times 108 - 26^{2}] [7 \times 87 - 23^{2}]}} = 0.4875$ $and the vector cosine distance is \frac{91}{\sqrt{108} \cdot \sqrt{87}} = 0.938793$
Another example of the item similarity criteria is a specified number of co-ratings between items. A co-rating exists for two items when a user rates both items. Thus, the item similarity criteria may specify a minimum number of co-ratings that must be shared between two items for the two items to be considered similar items. Yet another example of the item similarity criteria is a specified fraction of co-ratings between items divided by the number of ratings of either item. Thus, the item similarity criteria may specify a minimum fraction of co-ratings between two items divided by the number of ratings of either item that must be satisfied for the two items to be considered similar items. If the total population of users is U, the number of users who rated item A is R_a, the number of users who rated item B is R_b, and if the probability of a user rating item A is independent of the probability of the user rating item B, then the expected number of co-ratings E(R_ab) is given by: $E (R_{ab}) = U • \frac{R_{a}}{U} • \frac{R_{b}}{U}$
If the actual number of co-raters of both Item A and Item B R_abis much larger than E(R_ab), it is likely that Item A and Item B are related in some way. On the other hand, if R_abis much smaller than E(R_ab), it is likely that Item A and Item B are dissimilar in a way such that raters of Item A are unlikely to encounter Item B, or unlikely to have an opinion about Item B and so are less likely to rate Item B (and vice-versa). Either of these pieces of information may be used as additional similarity criteria. Note: if one of Item A or Item B is newer than the other (in other words, did not exist at the time the other Item was popular or well-known), then it is very likely that the total number of co-ratings of the Items will be much smaller than would be expected by the basic formula above. As a variation, the pool of ratings may be subdivided into slices by time—for example, months—and the co-raters fraction calculation described above may be applied to in each successive (month) slice.
In some situations, using the Pearson Correlation Coefficient to measure similarity between two items can make two dissimilar items appear similar. This occurs where the underlying rating distributions of two items make the co-ratings of the two items appear to be correlated. For example, two items may have a high number of high ratings. Although the items are unrelated, a grouping of rating pairs in both items among the co-raters, i.e., users who rated both items, makes them appear related.
According to one embodiment of the invention, the ratings distributions of the two items are “factored out.” This is performed by taking the percentage of each rating for each item, and multiplying it out in a two-dimensional array to obtain expected percentages for every possible pair of ratings (r1, r2) for the two items. Multiplying each of the array values by the total number of co-raters produces the expected count of co-raters for each rating pair, if the items are unrelated. The actual counts of co-ratings in each of the array cells are compared to see if they are more or less than expected. Higher than expected values in areas corresponding to similarity, e.g. a user that rated both items high or both item slow, indicates that the items are similar. Higher than expected co-rater counts in areas where a user liked one item and disliked another indicates that the items are not positively related. This use of the underlying ratings distribution for all raters provides a better understanding of values for co-raters.
The descriptions above have identified several different similarity criteria, which may be combined to determine overall similarity. For example, in one embodiment, the Pearson Correlation Coefficient is required to be greater than a threshold (e.g. 0.3), and also the fraction of co-raters R_abmust be at least some fraction (e.g. one half) the expected number of co-raters E(Rab), and the absolute number of co-raters must be at least some minimum number (e.g. 200).
The methods described above show how to calculate the similarity between two items in order to find a set of similar items. The exact same methods may be applied in order to calculate the similarity between two users in order to find a set of similar users.
In some circumstances, there may be insufficient rating data to determine a set of items similar to the particular item. For example, if the particular item is new, there may be very little data allowing correlations to be calculated. In this case, other semantic information may be used to help determine similar items.
In one embodiment, items are additionally classified by semantic feature values. For example, if the items are movies, the semantic features might be genre (or category), director, lead actors/actresses, other actors/actresses, etc., and the feature values are the genre, director, or actor names. In a music domain, the features might be styles, soloists, musicians, and lyricists. Each type of feature is given an importance score, which may be positive, negative, or zero, and two items may be compared by counting the number of feature matches and summing their importance scores. For example, if a movie director has an importance score of 3, and a lead actor has an importance score of 1, then a pair of movies sharing one director and one lead actor would score a similarity of 4, while a different pair of movies sharing two lead actors would have a similarity score of 2. The absence of matches may be accorded a negative score (for example, the absence of matching directors might score the pairing −0.1 points). The set of similar items is the specified number of items whose similarity score with the particular item is largest, or the set of items whose similarity score is above a specified threshold.
In another embodiment, tools may be provided for domain experts to specify item similarity by non-machine-implemented means. Once additional ratings data becomes available, the similar items specified by domain experts may be supplemented or discarded.
C. Identifying Other Users that have Given Ratings for the Other Items Similar to the Particular User
Once similar items 812 have been determined, one or more users from users 804 that have given ratings to similar items 812 that are similar to ratings given by the particular user 808 to similar items 812 are identified. All of users 804 may be evaluated for this purpose. There may be a significant and unacceptable computational cost, however, in evaluating all of users 804 in some applications where the number of users 804 is large. Furthermore, some users may be less useful than others because they have very little ratings data or have ratings data with poor statistical characteristics, as described in more detail hereinafter.
According to one embodiment of the invention, a subset of users 804, referred to herein as reference users 814, is determined based upon specified user selection criteria. Reference users 814 are users from all users 804 that satisfy the specified user selection criteria. A wide variety of user selection criteria may be used, depending upon the requirements of a particular application, and the invention is not limited to any particular user selection criteria. According to one embodiment of the invention, the user selection criteria include a minimum number of ratings. Thus, users from users 804 that do not have at least the minimum number of ratings are not included in reference users 814. The user selection criteria may include a maximum number of ratings. Thus, users from users 804 that have more than the maximum number of ratings are not included in reference users 814.
According to another embodiment of the invention, the user selection criteria specify an average rating range. Thus, users from users 804 with an average rating that is outside the specified range are not included in reference users 814. For example, suppose that ratings are provided on a numerical scale from one to five. Suppose further that the user selection criteria specify that the acceptable average rating range is two to four. In this situation, users 804 with an average rating that is less than two or greater than four are not included in reference users 814.
According to another embodiment of the invention, the user selection criteria specify that user ratings must conform to within a specified tolerance of a Normal distribution. Thus, users from users 804 with a rating distribution that does not approximate a Normal distribution within a specified tolerance are not included in reference users 814. For example, suppose that ratings are on a whole number numerical scale of one to five. A user that has more one ratings than two ratings, more two ratings than three ratings, more five ratings than four ratings, or more four ratings than three ratings is not included in reference users 814. The specified tolerance may vary, depending upon the requirements of a particular application. For example, some applications may use a specified tolerance that allows a user in reference users 814 to have a slightly greater number of two ratings than three ratings, and an equal number of two ratings and one ratings, but not more one ratings than three ratings.
According to another embodiment of the invention, the user selection criteria specify a range of acceptable standard deviations of user ratings. Thus, users from users 804 with a standard deviation of ratings that is not within the range of acceptable range of standard deviations of ratings are not included in reference users 814. For example, the user selection criteria might specify a standard deviation range of between 10% and 20%. Thus, a user having a standard deviation of ratings of less than 10% or 20% or greater is not included in reference users 814.
D. Generate Estimation of how the Particular User Would Rate the Particular Item
Generating an estimation of how particular user 808 would rate particular item 806 generally involves identifying which users 804 have rated similar items 812 items similarly to particular user 808, and then generating the estimation based upon how those users 804 rated the particular item 806. In the situation where reference users 814 are identified from users 804 as described herein, then the approach generally involves identifying which reference users 814 have rated similar items 812 items similarly to particular user 808, and then generating the estimation based upon how those reference users 814 rated the particular item 806. Embodiments of generating an estimation of how particular user 808 would rate particular item 806 are described herein in the context of using reference users 814, although the invention is not limited to this context.
According to one embodiment of the invention, reference users' 814 ratings of similar items 812 are compared to the particular user's 808 ratings of similar items 812 to identify reference users 814 that are the most similar to particular user 808. In FIG. 8, reference users 814 that have provided ratings for similar items 812 that are most similar to the ratings for similar items 812 provided by particular user 808 are referred to hereinafter as similar reference users 816. In FIG. 8, the horizontal lines depicted across similar items 812 for the similar reference users 816 are provided for illustration purposes only and are not meant to imply that the similar reference users 816 have necessarily provided ratings for all similar items 812. Similarly, it is not necessary for the particular user 808 to have rated all similar items 812.
According to one embodiment of the invention, rating similarity criteria are used to identify the similar reference users 816. The rating similarity criteria may be similar to the item similarity criteria described herein. For example, the similar reference users 816 may be determined based upon determining statistical correlations between the ratings of similar items 812 provided by reference users 814 and the particular user 808 and then selecting the similar reference users 816 based upon the reference users 814 that best correlate to the particular user 808. The similarity between similar items 812 and the particular item 806 may also be considered in identifying the similar reference users 816. More specifically, ratings of similar items 812 that are more similar to the particular item 806 may be given more weight in determining the similar reference users 816 from reference users 814.
Once the similar reference users 816 are identified, the estimation of how the particular user 808 would rate the particular item 806 is generated based upon the ratings of the particular item 806 made by the similar reference users 816. The degree of similarity between the similar reference users 816 and the particular user 808 may vary based upon a variety of factors including, for example, the number of similar items 812 that both the particular user 808 and the similar reference users 816 have rated and how the particular user 808 and the similar reference users 816 rated the similar items 812. The estimation may be generated in different ways, depending upon the similarity between the ratings of similar items 812 provided by the similar reference users 816 and the particular user 808. For example, there may be situations where the ratings provided by the one or more similar reference users 816 for similar items 812 are identical to the ratings provided by the particular user 808 for the same similar items 812. This may occur, for example, where the number of similar items 812 rated by both the similar reference users 816 and the particular user 808 is relatively small. In this situation, the estimated rating of the particular item 806 by the particular user 808 may be the average rating made by the one or more similar reference users 816 for the particular item 806.
In situations where the ratings for similar items 812 by the similar reference users 816 and the particular user 808 are not identical, then the estimated rating of the particular item 806 by the particular user 808 may be the weighted linear sum of the ratings made by the one or more similar reference users 816 for the particular item 806. The weighting applied varies based upon the similarity between each similar reference users 816 and the particular user 808. According to one embodiment of the invention, the weighted linear sum of the ratings for particular item 806 is the sum of the products of the weights and the similar reference users' 816 ratings of particular item 806, divided by the sum of the weights. Thus, the ratings of particular item 806 provided by the similar reference users 816 that are most similar to the particular user 808 are given more weight than the ratings of particular item 806 provided by the similar reference users 816 that are less similar to the particular user 808. For example, suppose that the similar reference users 816 include users A and B that are determined to have similarities to the particular user 808 of 0.8 and 0.5 (on a scale of 0 to 1), respectively. Furthermore, users A and B have given ratings for particular item 806 of 5 and 3 (on a scale of 1 to 5), respectively. The average rating of particular item 806 by users A and B is 4 ((5+3)/2). The weighted linear sum of the ratings for particular item 806 by users A and B is 4.23 ((5*0.8+3*0.5)/1.3). The weighted linear sum of the ratings is greater than the average of the ratings in this situation because the higher rating of user A is given greater weight than the lower weighting of user B, by virtue of the relatively greater similarity between user A and the particular user 808 than between user B and the particular user 808.
Various confidence metrics may be calculated that indicate a confidence level in an estimated rating of particular item 806 by the particular user 808. According to one embodiment of the invention, a confidence metric is calculated based upon the disparity or spread between estimated ratings, where each estimated rating is based upon a different one of the similar reference users 816. The standard deviation of the estimated ratings may be used for this purpose. The calculation of confidence based upon the disparity between estimated ratings may weight these estimated ratings based on the strength of the similarity between the reference user 816 and the particular user 808. A confidence metric may also be made based upon similarities between the particular user 808 and each similar reference user 816 that is used to estimate a rating. The calculation of confidence may also take into account the absolute strength of the similarity between the reference users 816 used to generate the prediction and the particular user 808.
There may be situations where there is an insufficient (or zero) number of similar reference users 816. In these situations, a weighted average of all reference users 814, who each share at least a rating of one of similar items 812 with the particular user 808 and who have each rated the particular item 806, may be used to generate an estimation of how the particular user 808 would rate the particular item 806. According to one embodiment of the invention, the weightings are applied based upon the similarity of each reference user 814 to the particular user 808. Confidence metrics may also be applied to generating an estimation of how the particular user 808 would rate the particular item 806 based upon all reference users 814 used to generate the estimate.
In situations where there is an insufficient (or zero) number of reference users 814, which may also occur if the particular user 808 has an insufficient (or zero) number of ratings, an average rating of the particular item 806 by all reference users 814 who have rated the particular item 806 may be used as an estimate of how the particular user 808 would rate the particular item 806. Alternatively, an average rating of all users 804 who have rated the particular item 806 may be used as an estimate of how the particular user 808 would rate the particular item 806. In these situations, a weighted average of all reference users 814, who each share at least a rating of one of similar items 812 with the particular user 808 and who have each rated the particular item 806 may be used.
E. Recommendation Architecture
FIG. 9 is a block diagram that depicts an arrangement 900 for recommending items according to an embodiment of the invention. Arrangement 900 includes a Web server 902 communicatively coupled to a recommendation system 904. As described in more detail hereinafter, recommendation system 904 is configured to process ratings data using prioritized tasks to provide improved availability and accuracy of item recommendations. Recommendation system 904 includes a recommendation engine 906, a non-volatile storage 908 and a ratings database (DB) 910. Recommendation engine 906 may be implemented by any mechanism or any combination of hardware and software for performing the functions described herein. For purposes of explanation only, recommendation engine 906 is described in the context of a software process executing in a volatile memory. Non-volatile storage 908 may be implemented by any type of non-volatile storage mechanism and the invention is not limited to any particular mechanism. Example implementations include, without limitation, non-volatile memory and disc storage. Ratings DB 910 may be implemented by any database management mechanism and the invention is not limited to any particular implementation.
Recommendation engine 906 maintains ratings and prediction data 912 that includes users' actual ratings and estimated ratings calculated according to the algorithms described herein, or otherwise. Ratings and prediction data 912 may be maintained in a volatile memory, such as a random access memory (RAM), to reduce access times. According to one embodiment of the invention, ratings and prediction data 912 is maintained in a RAM array and is addressed using 64 bit addressing as provided by high-end computing architectures. The array is laid out such that two bytes are used for each cell, containing a rating value, an estimated rating value, a confidence value, and one or more flags indicating the presence and validity of the values. The RAM array may be considered as divided into a number of pages of perhaps 8,192 bytes each (which corresponds to the physical structure of current computers). The cells are laid out so that each user's ratings and estimated ratings span one or more pages for locality of access.
According to one embodiment of the invention, ratings and prediction data 912 is periodically saved to a data file 914 maintained on non-volatile storage 908. For example, low-level operating system primitives may be used to ensure that data written to the RAM array is also stored in data file 914 in a timely manner, so that recommendation engine 906 may be periodically shut down for maintenance and upgrade without losing any of the ratings and prediction data 912. Upon restarting the recommendation engine 906, data from data file 914 is mapped back (reloaded) into the RAM array, thereby restoring all the results of previous computations. In one embodiment of the invention, the memory mapping functionality of the Solaris operating system offered by Sun Microsystems, Inc., permits data file 914 to be sparsely populated with disc blocks, such that any pages in the RAM array that have no values do not occupy space on non-volatile storage 908. This provides considerable savings in hardware cost, since much of the RAM array will never contain ratings and prediction data. Ratings data is also stored in ratings DB 910, which allows data file 914 to be re-generated in the event of a failure of non-volatile storage. Ratings DB 910 may also store any data contained in the RAM array or data file 914 or any other data used by recommendation system 904, depending upon the requirements of a particular implementation.
The use of data file 914 also allows storage of more ratings and prediction data 912 than can be stored in the RAM array, given practical limitations in available RAM on computer architectures. According to one embodiment of the invention, the size of the RAM array is selected to be large enough to store ratings and prediction data 912 for at least reference users 814 and active users of recommendation system 904. As users' status changes from active to inactive, the physical pages of storage in the RAM array allocated to their rating and prediction data 912 may be reclaimed and reused for other users just becoming active.
Recommendation engine 906 is configured to perform a wide variety of tasks. These tasks may be prioritized in one or more task queues 916 to coordinate the activity of one or more processing threads. The one or more task queues may be stored in a volatile storage, such as RAM, or in non-volatile storage 908. The processing threads process tasks from the task queues 916 based upon priority and the processing of a task may create additional queue requests that are added to task queues 916. For example, recommendation engine 906 is configured to process requests for estimated ratings and requests for recommendations. An estimated rating for a particular item for a particular user may be requested in situations where, for example, a user is viewing data associated with the particular item, such as a review, on a graphical user interface (GUI), and it is desirable to display data that indicates an estimate of how the particular user would rate the particular item. A request for recommendations may be made, for example, in situations where a user has requested a list of recommended items from a Web application. When a request for an estimated rating a particular item for a particular user is received from Web server 902, recommendation engine determines whether the requested estimated rating has previously been calculated and stored in the RAM array. If so, then the estimated rating is immediately returned to Web server 902. If not, then a request is queued on the task queue with high priority to calculate the estimated rating (according to the algorithms described herein, or otherwise). If the compute threads complete the calculation of the estimated rating within some small threshold of time (for example, half a second) then the calculated estimated rating is returned. Otherwise, the recommendation engine 906 is configured to provide an average rating of the particular item as the estimated rating. When a request for a recommendation for a specific user is received from Web server 902, recommendation engine 906 identifies the set of items with the highest estimated ratings and highest confidence for that user and returns to Web server 902 data that identifies the set of items as recommendations. If there are no confident predictions, then no recommendations may be returned. Although embodiments of the invention are described herein and depicted in the figures in the context of a single Web server 902, this is provided for purposes of explanation only. Recommendation system 904 may be used with any number of Web servers and also other types of computing entities.
Recommendation engine 906 is also configured to receive and process rating data from Web server 902. Upon receipt of rating data, which may include a new or changed rating, recommendation engine 906 stores the rating in the RAM array using a simple address calculation and a simple memory write. The rating is also queued in a task queue 916 to be written to ratings DB 910. In addition, estimated ratings for items similar to the item associated with the received rating data may need to be updated to reflect the received rating data. Therefore, one or more tasks are added to task queues 916 to recalculate the estimated ratings for items for which the item associated with the received rating data is considered similar. According to one embodiment of the invention, these tasks are given high priority, since the updated estimated ratings may be required in a relatively short time, for example for a merchandising decision or a recommendation for a user who is presumably visiting the Website in order to provide the rating. Quality metrics may be calculated from comparisons of actual user ratings to estimated user ratings. For example, suppose that the estimated rating 810 has been determined for the particular item 806 and the particular user 808. Suppose further that this estimated rating 810 has been provided to Web server 902, displayed for the particular user 808, and the particular user 808 has provided an actual rating for the particular item 806. A quality metric may be calculated based upon a comparison of the actual rating to the estimated rating 810. This quality metric, along with other quality metrics, may be used to fine tune future estimated ratings.
Recommendation engine 906 is further configured to periodically, and continuously if necessary, recalculate ratings and prediction data 912. In particular, recommendation engine 906 is configured to recalculate the set of similar items for each item to reflect changes in ratings data and new ratings data. Changes in ratings data and new ratings data may be tracked and various thresholds and aging schemes used to prioritize recalculation of similar item sets. For example, items that have a relatively large amount of ratings data updated may be given priority over an item for which the calculated similar item set is simply old. It is beneficial to periodically recalculate the set of similar items for a particular item, even if no ratings data has changed for the particular item, since ratings data may have changed the set of items considered similar to the particular item.
There may be situations where limitations in the power of the computing architecture used to implement recommendation system 904 make it difficult for recommendation system 904 to keep up with changes in ratings data. That is, the estimated ratings maintained by recommendation engine 906 may fall behind the updated ratings data. One consequence of this situation may be a loss in accuracy of estimated ratings. There may also be situations where limitations in the available size of non-volatile storage 908 may make maintaining data file 914 difficult. Accordingly, according to another embodiment of the invention, the approach is segmented into a series of separate recommendation engines and separate disc arrays allowing scaling to much higher capacities and load levels. FIG. 10 depicts an example embodiment of such an arrangement 1000.
Arrangement 1000 includes a Web server 1002 and a recommendation system 1004, configured in accordance with an embodiment of the invention. Recommendation system 1004 includes a similars/reference engine 1006, a non-volatile storage 1008, a ratings DB 1010 and recommendation engines 1012, 1014.
A single reference users/similar items data file 1016 is used to store the ratings for reference users 814. Reference users/similar items data file 1016 is maintained on non-volatile storage 1008 by similars/reference engine 1006, which maintains the set of reference users 814. Reference users 814 are chosen as described herein. If too many of users 804 qualify as reference users 814, then users 804 who have the most attractive rating patterns are chosen as reference users 814. Example criteria for choosing the reference users 814 include, without limitation, users with the most ratings, or the most ratings on otherwise unrated items, or other criteria that may be appropriate. New candidate reference users are loaded from ratings DB 1010 as they qualify, and existing reference users 814 may be displaced if any of the candidate reference users have more attractive parameters making them more suitable as references users 814 for making estimated ratings. Similars/reference engine 1006 is also configured to calculate sets of similar items 812 for each of items 802 on a continuous basis and store that data into reference users/similar items data file 1016.
Recommendation engines 1012, 1014 are configured to generate estimates of how users would rate items, as previously described herein. Recommendation engines 1012, 1014 have exclusive read/write access to active user data files 1018 and 1020, respectively. Active user data files 1018, 1020 store ratings and prediction data for the users assigned to recommendation engines 1012, 1014. Recommendation engines 1012, 1014 also have read-only access to reference users/similar items data file 1016. Read-only access is inexpensive in resources on the disc system and computer IO systems except when the recommendation engines 1012, 1014 are being started up. In an alternative embodiment, the reference users/similar items data file 1016 is periodically copied one or more times, and the resulting copy or copies are mapped by the recommendation engines 1012, 1024. Recommendation engines 1012, 1014 estimate how users would rate items as described herein, obtaining the reference rating values from reference users/similar items data file 1016 and storing estimated ratings in their private read-write active user data files 1018, 1020. The assignment of users to recommendation engines 1012, 1014 may be made using a variety of techniques, depending upon the requirements of a particular implementation, and the invention is not limited to any particular approach or assignment of users. Furthermore, the assignment of users to recommendation engines 1012, 1014 may be dynamic and change over time, depending, for example, on loading requirements.
Reference Users' 814 active profiles of ratings and predictions are stored in one of the active user data files 1018, 1020 and estimated ratings for reference users 814 are calculated by the appropriate recommendation engine 1012, 1014 and stored in the corresponding active user data files 1018, 1020. When a reference user 814 updates their ratings, those ratings are also stored in ratings DB 1010. Sometime later, similars/reference engine 1006 updates the shared copy of the ratings for that reference user 814 in reference user/similar items data file 1016. As long as the rate of change of ratings for reference users 814 is small compared to the total number of ratings for reference users 814, the quality of the similar items and estimated ratings calculated using the older data will not suffer measurably. Recommendation engines 1012, 1014 may store ratings and prediction data in one or more local RAM arrays and then periodically write the ratings and prediction data to active user data files 1018, 1020.
Task queues may be shared between the one or more engines, so that new work items may be triggered by one recommendation engine, but performed by a different engine if the task relates to data managed by the other engine.
Although embodiments of the invention have been described herein in the context of using a single similars/reference engine 1006, additional scaling may be provided by using multiple similars/reference engines that are each assigned responsibility for a portion of reference users 814. Since the determination of similar items 812 requires comparing columns of ratings for pairs of items across all reference users 814, it is appropriate to split reference users 814 by vertical segmentation by items, for example such that even/odd items (or other item segmentation) are split between different similars/reference engines. In this situation, each portion of the ratings data for each reference user 814 is available to each similars/reference engine. More specifically, each data file is owned by a single similars/reference engine that has read/write access for maintenance, and each data file is also made available on a read-only basis to all other similars/reference engine for access to the reference ratings data.
VIII. Implementation Mechanisms
The approach described herein for estimating how a user would rate a particular item is applicable to any type of items or content and is well suited for Internet-based rental applications for renting movies, music and games to customers. The invention may be implemented in hardware circuitry, in computer software, or a combination of hardware circuitry and computer software and is not limited to a particular hardware or software implementation. The invention is ideally suited for recommendation engine applications and may be implemented as a stand-alone mechanism, or integrated into a rental mechanism. For example, the approach is ideally suited for identifying rental items, such as movies, music and games, to recommend to users.
FIG. 11 is a block diagram that illustrates a computer system 1100 upon which an embodiment of the invention may be implemented. Computer system 1100 includes a bus 1102 or other communication mechanism for communicating information, and a processor 1104 coupled with bus 1102 for processing information. Computer system 1100 also includes a main memory 1106, such as a random access memory (RAM) or other dynamic storage device, coupled to bus 1102 for storing information and instructions to be executed by processor 1104. Main memory 1106 also may be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 1104. Computer system 1100 further includes a read only memory (ROM) 1108 or other static storage device coupled to bus 1102 for storing static information and instructions for processor 1104. A storage device 1110, such as a magnetic disk or optical disk, is provided and coupled to bus 1102 for storing information and instructions.
Computer system 1100 may be coupled via bus 1102 to a display 1112, such as a cathode ray tube (CRT), for displaying information to a computer user. An input device 1114, including alphanumeric and other keys, is coupled to bus 1102 for communicating information and command selections to processor 1104. Another type of user input device is cursor control 1116, such as a mouse, a trackball, or cursor direction keys for communicating direction information and command selections to processor 1104 and for controlling cursor movement on display 1112. This input device typically has two degrees of freedom in two axes, a first axis (e.g., x) and a second axis (e.g., y), that allows the device to specify positions in a plane.
The invention is related to the use of computer system 1100 for implementing the techniques described herein. According to one embodiment of the invention, those techniques are performed by computer system 1100 in response to processor 1104 executing one or more sequences of one or more instructions contained in main memory 1106. Such instructions may be read into main memory 1106 from another machine-readable medium, such as storage device 1110. Execution of the sequences of instructions contained in main memory 1106 causes processor 1104 to perform the process steps described herein. In alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions to implement the invention. Thus, embodiments of the invention are not limited to any specific combination of hardware circuitry and software.
The term “machine-readable medium” as used herein refers to any medium that participates in providing data that causes a machine to operation in a specific fashion. In an embodiment implemented using computer system 1100, various machine-readable media are involved, for example, in providing instructions to processor 1104 for execution. Such a medium may take many forms, including but not limited to, non-volatile media, volatile media, and transmission media. Non-volatile media includes, for example, optical or magnetic disks, such as storage device 1110. Volatile media includes dynamic memory, such as main memory 1106. Transmission media includes coaxial cables, copper wire and fiber optics, including the wires that comprise bus 1102. Transmission media can also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications.
Common forms of machine-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, or any other magnetic medium, a CD-ROM, any other optical medium, punchcards, papertape, any other physical medium with patterns of holes, a RAM, a PROM, and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave as described hereinafter, or any other medium from which a computer can read.
Various forms of machine-readable media may be involved in carrying one or more sequences of one or more instructions to processor 1104 for execution. For example, the instructions may initially be carried on a magnetic disk of a remote computer. The remote computer can load the instructions into its dynamic memory and send the instructions over a telephone line using a modem. A modem local to computer system 1100 can receive the data on the telephone line and use an infrared transmitter to convert the data to an infrared signal. An infrared detector can receive the data carried in the infrared signal and appropriate circuitry can place the data on bus 1102. Bus 1102 carries the data to main memory 1106, from which processor 1104 retrieves and executes the instructions. The instructions received by main memory 1106 may optionally be stored on storage device 1110 either before or after execution by processor 1104.
Computer system 1100 also includes a communication interface 1118 coupled to bus 1102. Communication interface 1118 provides a two-way data communication coupling to a network link 1120 that is connected to a local network 1122. For example, communication interface 1118 may be an integrated services digital network (ISDN) card or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, communication interface 1118 may be a local area network (LAN) card to provide a data communication connection to a compatible LAN. Wireless links may also be implemented. In any such implementation, communication interface 1118 sends and receives electrical, electromagnetic or optical signals that carry digital data streams representing various types of information.
Network link 1120 typically provides data communication through one or more networks to other data devices. For example, network link 1120 may provide a connection through local network 1122 to a host computer 1124 or to data equipment operated by an Internet Service Provider (ISP) 1126. ISP 1126 in turn provides data communication services through the worldwide packet data communication network now commonly referred to as the “Internet” 1128. Local network 1122 and Internet 1128 both use electrical, electromagnetic or optical signals that carry digital data streams. The signals through the various networks and the signals on network link 1120 and through communication interface 1118, which carry the digital data to and from computer system 1100, are exemplary forms of carrier waves transporting the information.
Computer system 1100 can send messages and receive data, including program code, through the network(s), network link 1120 and communication interface 1118. In the Internet example, a server 1130 might transmit a requested code for an application program through Internet 1128, ISP 1126, local network 1122 and communication interface 1118. The received code may be executed by processor 1104 as it is received, and/or stored in storage device 1110, or other non-volatile storage for later execution. In this manner, computer system 1100 may obtain application code in the form of a carrier wave.
The novel approach described herein for renting items to customers allows the decision of what items to rent to be separated from the decision of when to rent the items. Customers may specify what items to rent using the item selection criteria and receive the items at a future point in time, without having to go to the provider to pick up the items. The selection criteria may be user specific and may indicate a desired fulfillment sequence. Furthermore, customers are not constrained by conventional due dates and instead may establish continuous, serialized rental streams of items. The approach also allows more efficient inventory management.
The “Max Out” approach for inventory management allows users to maintain their own inventory of items that are periodically replaced by other items according to specified event criteria. The event criteria that trigger sending another item to a customer are very flexible and may be tailored to the requirements of a particular application. For example, as described herein, the event criteria may include a return of any of the items currently in use by the customer or merely customer notification. This is very convenient in the context of movie rentals since the return of a movie to the provider automatically triggers the sending of another movie to the customer. The “Max Turns” approach for inventory management, when used alone or in combination with “Max Out,” provides even greater flexibility for customers and providers. The max number of turns can be selected individually for particular customers depending upon their particular needs.
The “Max Out” and “Max Turns” approaches provide great flexibility in establishing subscription models to satisfy the needs of a particular application. Specifically, the size and replacement frequency of customer inventories can be tailored to each customer with individualized subscription plans.
In the foregoing specification, the invention has been described as applicable to an implementation anticipating Internet based ordering and mail or other long-distance delivery of the items, where the special advantages of the method are very attractive. However the same invention may be applied in a more conventional video, games, or music rental-store setting, where subscription customers may be allowed rentals of a specified number of movies, games, or music selections at any time, and/or in one subscription period, without rental return due dates, in exchange for a periodic rental subscription fee.
In the foregoing specification, embodiments of the invention have been described with reference to numerous specific details that may vary from implementation to implementation. Thus, the sole and exclusive indicator of what is, and is intended-by the applicants to be, the invention is the set of claims that issue from this application, in the specific form in which such claims issue, including any subsequent correction. Hence, no limitation, element, property, feature, advantage or attribute that is not expressly recited in a claim should limit the scope of such claim in any way. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense.

Claims

1. A machine-implemented method for estimating how a particular user would rate a particular item from a plurality of items, the method comprising the machine-implemented steps of:

identifying, from the plurality of items, one or more other items that have ratings similar to the particular item;

selecting a set of other users from a plurality of other users based upon user selection criteria;

identifying one or more other users from the set of other users that have given ratings to a subset of the one or more other items, wherein the ratings are statistically similar to ratings given by the particular user to the subset of the one or more other items; and

generating an estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users.

2. The machine implemented method as recited in claim 1, wherein the step of identifying, from the plurality of items, one or more other items that have ratings similar to the particular item is performed off-line in advance of the other steps.

3. A machine-readable medium for estimating how a particular user would rate a particular item from a plurality of items, the machine-readable medium carrying instructions which, when processed by one or more processors, cause performance of the steps of:

4. The machine-readable medium as recited in claim 3, wherein the step of identifying, from the plurality of items, one or more other items that have ratings similar to the particular item is performed off-line in advance of the other steps.

5. An apparatus for estimating how a particular user would rate a particular item from a plurality of items, the apparatus comprising a memory storing instructions which, when processed by one or more processors, cause performance of the steps of:

6. The apparatus as recited in claim 5, wherein the step of identifying, from the plurality of items, one or more other items that have ratings similar to the particular item is performed off-line in advance of the other steps.

7. A machine-implemented method for estimating how a particular user would rate a particular item from a plurality of items, the method comprising the machine-implemented steps of:

identifying one or more items from the plurality of items that have ratings similar to the particular item;

identifying one or more other users that have given ratings to the one or more items that are substantially similar to ratings given by the particular user to the one or more items; and

8. The machine implemented method as recited in claim 7, wherein the step of identifying, from the plurality of items, one or more other items that have ratings similar to the particular item is performed off-line in advance of the other steps.

9. The machine-implemented method as recited in claim 7, wherein identifying one or more items from a plurality of items that have ratings similar to the particular item includes calculating a statistical correlation between the particular item and the other items from the plurality of items.

10. The machine-implemented method as recited in claim 7, further comprising the machine-implemented step of determining whether the particular item and each of the one or more items have at least a specified number of co-ratings.

11. The machine-implemented method as recited in claim 7, further comprising the machine-implemented step of determining whether the particular item and one of the one or more items have at least a specified fraction of co-ratings between the particular item and the one of the one or more items divided by the number of ratings of either the particular item or the one of the one or more items.

12. The machine-implemented method as recited in claim 7, wherein the step of identifying one or more items from the plurality of items that have ratings similar to the particular item includes identifying whether the particular item and the one or more items have a number of co-raters greater than a specified fraction of a statistically expected number, given the number of ratings of each item counted separately.

13. The machine-implemented method as recited in claim 7, wherein the step of identifying one or more items from the plurality of items is performed by one or more external agents.

14. The machine-implemented method as recited in claim 7, wherein the step of identifying one or more items from the plurality of items that have ratings similar to the particular item includes consideration of one or more of a statistical correlation between the particular item and the one or more items, whether the particular item and the one or more items have at least a specified number of co-ratings, whether the particular item and the one or more items have at least a specified fraction of co-ratings, and whether the particular item and the one or more items have at least a specified fraction of the statistically expected number of co-ratings.

15. The machine-implemented method as recited in claim 7, further comprising the machine-implemented step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having at least a specified minimum number of ratings.

16. The machine-implemented method-as recited in claim 7, further comprising the machine-implemented step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having less than a specified maximum number of ratings.

17. The machine-implemented method as recited in claim 7, further comprising the machine-implemented step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having an average rating within a specified range.

18. The machine-implemented method as recited in claim 7, further comprising the machine-implemented step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having a rating distribution within a specified tolerance of a Normal distribution.

19. The machine-implemented method as recited in claim 7, further comprising the machine-implemented step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having a rating standard deviation within a specified range.

20. The machine-implemented method as recited in claim 7, wherein identifying one or more other users that have given ratings to the one or more items that are substantially similar to ratings given by the particular user to the one or more items includes consideration of one or more of whether the one or more other users have at least a specified minimum number of co-ratings, whether the one or more other users have less than a specified maximum number of ratings, whether the one or more other users have an average rating within a specified range, whether the one or more other users have a rating distribution within a specified tolerance of a Normal distribution, and whether the one or more other users have a rating standard deviation within a specified range.

21. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing the one or more other users for whom a statistical correlation with the particular user of co-ratings of the one or more items is greater than a specified threshold.

22. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing the one or more other users for whom a vector cosine distance of co-ratings of the one or more items is smaller than a specified threshold.

23. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing the one or more other users who have co-rated a subset of the one or more items exactly the same as the particular user.

24. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing the one or more other users who have rated co-rated a subset of the one or more items within a specified threshold of the ratings by the particular user.

25. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes consideration of one or more of whether the one or more users have a correlation of co-ratings of the one or more items greater than a specified threshold, whether the one or more users have a vector cosine distance of co-ratings of the one or more items smaller than a specified threshold, whether the one or more users have co-rated a subset of the one or more items exactly the same, and whether the one or more users having co-rated a subset of the one or more items within a specified threshold of the ratings by the particular user.

26. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing a specified number of other users most similar to the particular user.

27. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing other users whose similarity to the particular user is better than a specified threshold.

28. The machine-implemented method as recited in claim 7, wherein identifying the one or more other users includes choosing up to a specified number of other users most similar to the particular user, where all the chosen other users have similarity better than a specified threshold.

29. The machine-implemented method as recited in claim 7, wherein generating an estimation of how the user would rate the particular item based upon ratings for the particular item given by the one or more other users includes determining an average rating of the particular item by the one or more other users.

30. The machine-implemented method as recited in claim 7, wherein generating an estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users includes determining an average rating of the particular item by the one or more other users.

31. The machine-implemented method as recited in claim 7, further comprising the machine-implemented steps of:

generating, for each of the one or more users, a separate estimation of how the particular user would rate the particular item based upon the rating for the particular item given by each of the one or more other users, and

calculating a confidence metric based upon disparities between the separate estimations.

32. The machine-implemented method as recited in claim 7, wherein the plurality of items is a plurality of movies, the one or more items is one or more movies and the particular item is a particular movie.

33. The machine-implemented method as recited in claim 7, wherein the plurality of items is a plurality of music items, the one or more items is one or more music items and the particular item is a particular music item.

34. A machine-readable medium for estimating how a particular user would rate a particular item from a plurality of items, the machine-readable medium carrying instructions which, when processed by one or more processors, cause performance of the steps of:

identifying, from the plurality of items, one or more items that have ratings similar to the particular item;

35. The machine-readable medium as recited in claim 34, wherein the step of identifying, from the plurality of items, one or more other items that have ratings similar to the particular item is performed off-line in advance of the other steps.

36. The machine-readable medium as recited in claim 34, wherein identifying one or more items from a plurality of items that have ratings similar to the particular item includes calculating a statistical correlation between the particular item and the other items from the plurality of items.

37. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of determining whether the particular item and each of the one or more items have at least a specified number of co-ratings.

38. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of determining whether the particular item and one of the one or more items have at least a specified fraction of co-ratings between the particular item and the one of the one or more items divided by the number of ratings of either the particular item or the one of the one or more items.

39. The machine-readable medium as recited in claim 34, wherein the step of identifying one or more items from the plurality of items that have ratings similar to the particular item includes identifying whether the particular item and the one or more items have a number of co-raters greater than a specified fraction of a statistically expected number, given the number of ratings of each item counted separately.

40. The machine-readable medium as recited in claim 34, wherein the step of identifying one or more items from the plurality of items is performed by one or more external agents.

41. The machine-readable medium as recited in claim 34, wherein the step of identifying one or more items from the plurality of items that have ratings similar to the particular item includes consideration of one or more of a statistical correlation between the particular item and the one or more items, whether the particular item and the one or more items have at least a specified number of co-ratings, whether the particular item and the one or more items have at least a specified fraction of co-ratings, and whether the particular item and the one or more items have at least a specified fraction of the statistically expected number of co-ratings.

42. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having at least a minimum number of ratings.

43. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having less than a maximum number of ratings.

44. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having an average rating within a specified range.

45. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having a rating distribution within a specified tolerance of a Normal distribution.

46. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having a rating standard deviation within a specified range.

47. The machine-implemented method as recited in claim 34, wherein identifying one or more other users that have given ratings to the one or more items that are substantially similar to ratings given by the particular user to the one or more items includes consideration of one or more of whether the one or more other users have at least a specified minimum number of co-ratings, whether the one or more other users have less than a specified maximum number of ratings, whether the one or more other users have an average rating within a specified range, whether the one or more other users have a rating distribution within a specified tolerance of a Normal distribution, and whether the one or more other users have a rating standard deviation within a specified range.

48. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing the one or more other users for whom a statistical correlation with the particular user of co-ratings of the one or more items is greater than a specified threshold.

49. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing the one or more other users for whom a vector cosine distance of co-ratings of the one or more items is smaller than a specified threshold.

50. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing the one or more other users who have co-rated a subset of the one or more items exactly the same as the particular user.

51. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing the one or more other users who have rated co-rated a subset of the one or more items within a specified threshold of the ratings by the particular user.

52. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes consideration of one or more of whether the one or more users have a correlation of co-ratings of the one or more items greater than a specified threshold, whether the one or more users have a vector cosine distance of co-ratings of the one or more items smaller than a specified threshold, whether the one or more users have co-rated a subset of the one or more items exactly the same, and whether the one or more users having co-rated a subset of the one or more items within a specified threshold of the ratings by the particular user.

53. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing a specified number of other users most similar to the particular user.

54. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing other users whose similarity to the particular user is better than a specified threshold.

55. The machine-implemented method as recited in claim 34, wherein identifying the one or more other users includes choosing up to a specified number of other users most similar to the particular user, where all the chosen other users have similarity better than a specified threshold.

56. The machine-readable medium as recited in claim 34, wherein generating an estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users includes determining an average rating of the particular item by the one or more other users.

57. The machine-readable medium as recited in claim 34, wherein generating an estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users includes determining an average rating of the particular item by the one or more other users.

58. The machine-readable medium as recited in claim 34, further comprising one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional steps of:

59. The machine-readable medium as recited in claim 34, wherein the plurality of items is a plurality of movies, the one or more items is one or more movies and the particular item is a particular movie.

60. The machine-readable medium as recited in claim 34, wherein the plurality of items is a plurality of music items, the one or more items is one or more music items and the particular item is a particular music item.

61. An apparatus for estimating how a particular user would rate a particular item from a plurality of items, the apparatus comprising a memory storing instructions which, when processed by one or more processors, cause the one or more processors to perform the steps of:

62. The apparatus as recited in claim 61, wherein the step of identifying, from the plurality of items, one or more other items that have ratings similar to the particular item is performed off-line in advance of the other steps.

63. The apparatus as recited in claim 61, wherein identifying one or more items from a plurality of items that have ratings similar to the particular item includes calculating a statistical correlation between the particular item and the other items from the plurality of items.

64. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of determining whether the particular item and each of the one or more items have at least a specified number of co-ratings.

65. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of determining whether the particular item and one of the one or more items have at least a specified fraction of co-ratings between the particular item and the one of the one or more items divided by the number of ratings of either the particular item or the one of the one or more items.

66. The apparatus as recited in claim 61, wherein the step of identifying one or more items from the plurality of items that have ratings similar to the particular item includes identifying whether the particular item and the one or more items have a number of co-raters greater than a specified fraction of a statistically expected number, given the number of ratings of each item counted separately.

67. The apparatus as recited in claim 61, wherein the step of identifying one or more items from the plurality of items is performed by one or more external agents.

68. The apparatus as recited in claim 61, wherein the step of identifying one or more items from the plurality of items that have ratings similar to the particular item includes consideration of one or more of a statistical correlation between the particular item and the one or more items, whether the particular item and the one or more items have at least a specified number of co-ratings, whether the particular item and the one or more items have at least a specified fraction of co-ratings, and whether the particular item and the one or more items have at least a specified fraction of the statistically expected number of co-ratings.

69. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having at least a minimum number of ratings.

70. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having less than a maximum number of ratings.

71. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having an average rating within a specified range.

72. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having a rating distribution within a specified tolerance of a Normal distribution.

73. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional step of selecting the one or more other users from a plurality of other users based upon each of the one or more other users having a rating standard deviation within a specified range.

74. The apparatus as recited in claim 61, wherein identifying one or more other users that have given ratings to the one or more items that are substantially similar to ratings given by the particular user to the one or more items includes consideration of one or more of whether the one or more other users have at least a specified minimum number of co-ratings, whether the one or more other users have less than a specified maximum number of ratings, whether the one or more other users have an average rating within a specified range, whether the one or more other users have a rating distribution within a specified tolerance of a Normal distribution, and whether the one or more other users have a rating standard deviation within a specified range.

75. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing the one or more other users for whom a statistical correlation with the particular user of co-ratings of the one or more items is greater than a specified threshold.

76. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing the one or more other users for whom a vector cosine distance of co-ratings of the one or more items is smaller than a specified threshold.

77. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing the one or more other users who have co-rated a subset of the one or more items exactly the same as the particular user.

78. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing the one or more other users who have rated co-rated a subset of the one or more items within a specified threshold of the ratings by the particular user.

79. The apparatus as recited in claim 61, wherein identifying the one or more other users includes consideration of one or more of whether the one or more users have a correlation of co-ratings of the one or more items greater than a specified threshold, whether the one or more users have a vector cosine distance of co-ratings of the one or more items smaller than a specified threshold, whether the one or more users have co-rated a subset of the one or more items exactly the same, and whether the one or more users having co-rated a subset of the one or more items within a specified threshold of the ratings by the particular user.

80. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing a specified number of other users most similar to the particular user.

81. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing other users whose similarity to the particular user is better than a specified threshold.

82. The apparatus as recited in claim 61, wherein identifying the one or more other users includes choosing up to a specified number of other users most similar to the particular user, where all the chosen other users have similarity better than a specified threshold.

83. The apparatus as recited in claim 61, wherein generating an estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users includes determining an average rating of the particular item by the one or more other users.

84. The apparatus as recited in claim 61, wherein generating an estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users includes determining an average rating of the particular item by the one or more other users.

85. The apparatus as recited in claim 61, wherein the memory further comprises one or more additional instructions which, when processed by the one or more processors, cause the one or more processors to perform the additional steps of:

86. The apparatus as recited in claim 61, wherein the plurality of items is a plurality of movies, the one or more items is one or more movies and the particular item is a particular movie.

87. The apparatus as recited in claim 61, wherein the plurality of items is a plurality of music items, the one or more items is one or more music items and the particular item is a particular music item.

88. An apparatus configured to:

generate an estimation of how a particular user would rate a particular item from a plurality of items by:

generating the estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more other users.

89. The apparatus as recited in claim 88, wherein the step of generating the estimation is performed in advance of the need for the estimation and data that indicates the estimation is stored in a memory.

90. The apparatus as recited in claim 89, wherein the apparatus is further configured to write data that indicates the estimation to a non-volatile storage.

91. The apparatus as recited in claim 90, wherein the non-volatile storage is one or more magnetic or solid-state disc systems, all or portions of which are memory-mapped into a random-access memory as needed.

92. The apparatus as recited in claim 90, wherein the apparatus is further configured to retrieve the data that indicates the estimation from the non-volatile storage and write the data into a volatile memory.

93. The apparatus as recited in claim 88, wherein the apparatus is further configured to in response to one or more ratings being updated or added for any of the one or more other items or the particular item:

identifying one or more other items from the plurality of items that have ratings similar to the particular item, wherein the one or more other items may include any of the one or more items;

identifying one or more users that have given ratings to the one or more other items that are substantially similar to ratings given by the particular user to the one or more other items;

generating a revised estimation of how the particular user would rate the particular item based upon ratings for the particular item given by the one or more users; and

storing data that indicates the revised estimation of how the particular user would rate the particular item.

94. The apparatus as recited in claim 88, wherein the apparatus is further configured to in response to one or more ratings being updated or added for any of the one or more other items or the particular item:

identifying one or more estimated ratings that are affected by the one or more ratings being updated or added;

generating one or more updated estimated ratings for the one or more estimated ratings that are affected by the one or more ratings being updated or added; and

storing data that indicates the one or more updated estimated ratings.

95. The apparatus as recited in claim 88, wherein the apparatus is further configured to periodically generate and update similarity data that identifies similarities between items from the plurality of items based upon ratings associated with the plurality of items.

96. The apparatus as recited in claim 88, wherein the apparatus is further configured to select the one or more other users from a plurality of users based upon selection criteria that includes whether a particular user has a specified number of ratings and to generate and store data that indicates the selected one or more other users.

97. The apparatus as recited in claim 88, further comprising one or more prioritized task queues from which tasks are assigned to be processed based upon a priority associated with each task and wherein tasks associated with updating estimated ratings for active users are assigned the highest priority.

98. The apparatus as recited in claim 88, wherein the plurality of items is a plurality of movies, the one or more items is one or more movies and the particular item is a particular movie.

99. The apparatus as recited in claim 88, wherein the plurality of items is a plurality of music items, the one or more items is one or more music items and the particular item is a particular music item.

100. An apparatus comprising:

a plurality of processes assigned to generate a plurality of estimated ratings of how users would rate items from a plurality of items, wherein generating an estimate of how a particular user would rate a particular item from the plurality of items is performed by:

101. The apparatus as recited in claim 100, further comprising one or more processes that are configured to generate and store ratings data that indicates estimated ratings, wherein each such process is not given access to ratings data generated and stored by other such processes.

102. The apparatus as recited in claim 101, further comprising one or more processes assigned to select reference users and identify similarities between items from the plurality of items based upon ratings associated with the plurality of items and to generate and store similarity data that indicates the identified similarities between items from the plurality of items, wherein the one or more processes have read/write access to the similarity and reference user data and the one or more other processes have read-only access to the similarity data.

103. The apparatus as recited in claim 102, wherein the similarity and reference user data is copied periodically and the one or more recommendation processes have access to the copied similarity and reference user data.

104. The apparatus as recited in claim 100, further comprising an additional process assigned to select the one or more other users from a plurality of users based upon selection criteria that includes whether a user has a specified number of ratings and to generate data that indicates the selected one or more other users, wherein the additional process has read/write access to the data and the plurality of processes have read-only access to the data.

105. The apparatus as recited in claim 100, wherein the plurality of items is a plurality of movies, the one or more items is one or more movies and the particular item is a particular movie.

106. The apparatus as recited in claim 100, wherein the plurality of items is a plurality of music items, the one or more items is one or more music items and the particular item is a particular music item.

107. A machine-implemented method for estimating how a particular user would rate a particular item from a plurality of items, the method comprising the machine-implemented steps of:

identifying one or more items from the plurality of items that are similar to the particular item by associating semantic feature values with the plurality of items and the particular item, where each feature type has a match-score and a mismatch-score, either of which may be positive, zero, or negative, and scoring the similarity of items is performed by

finding the features that have matching values and summing the associated match-score, and

finding the features that have mismatched values and summing the associated mismatch-score,

wherein items with a summed similarity score above a specified threshold are considered similar or the specified number of items with the highest similarity score are considered similar;

108. A machine-readable medium for estimating how a particular user would rate a particular item from a plurality of items, the machine-readable medium carrying instructions which, when processed by one or more processors, cause performance of the steps of:

109. An apparatus for estimating how a particular user would rate a particular item from a plurality of items, the apparatus comprising a memory storing instructions which, when processed by one or more processors, cause performance of the machine-implemented steps of: